决策树部分理论支撑
1* 通过选取一定的特征来降低数据的不确定性（熵）
2* 建议寻找多分类问题的最优特征的最优候选值。把多分类问题转换成多几层递归的二分类问题，防止数据对特征值的控制敏感。
3* 停止条件
- 取得了最够好的分类结果
- 递归到了预定的最深深度
- 叶子节点的纯度
- 分裂次数达到极限
- 最大特征数
- . . .
4* 相关公式
-
e n t r o p y ( D ) = − ∑ i = 1 n P i l o g 2 P i entropy(D) = -\sum_{i=1}^n P_ilog_2 P_i
e
n
t
r
o
p
y
(
D
)
=
−
∑
i
=
1
n
P
i
l
o
g
2
P
i
e n t r o p y ( D , A ) = ∑ i = 1 k D A i D l o g 2 D A i entropy(D,A) = \sum_{i=1}^k \frac {D_{A_i}}{D} log_2D_{A_i} e n t r o p y ( D , A ) = ∑ i = 1 k D D A i l o g 2 D A i
g a i n ( D , A ) = e n t r o p y ( D ) − e n t r o p y ( D , A ) gain(D,A) = entropy(D) - entropy(D,A) g a i n ( D , A ) = e n t r o p y ( D ) − e n t r o p y ( D , A )
原本的熵 减去 考虑某种特征条件A之后的熵，得到信息增益
g a i n r a t e ( D , A ) = g a i n ( D , A ) / e n t r o p y ( D , A ) gain_rate(D,A) = gain(D,A)/entropy(D,A) g a i n r a t e ( D , A ) = g a i n ( D , A ) / e n t r o p y ( D , A )
同理，根据同样的方法可以得到 信息增益率
import
pandas
as
pd
import
numpy
as
np
df
=
pd
.
read_csv
(
'C:\\Users\\76485\\Desktop\\column.2C.csv'
)
'''
导入数据源&导入基本包;
交互命令窗口输入df.head()查看前五行数据
'''
X
=
df
.
drop
(
'V7'
,
axis
=
1
)
#drop进行有选择的数据删除,删除头标签为'V7'的列数据
y
=
df
.
V7
'''
轴用来为超过一维的数组定义属性，二维数据：0轴沿着行的方向向下，1轴沿着列的水平方向延伸
'''
'''
#将文字转换为数字的标准化程序
from sklearn.preprocessing import LabelEncoder
labelencoder = LabelEncoder()
for col in data.columns:
data[col] = labelencoder.fit_transform(data[col])
'''
#from sklearn.cross_validation import train_test_split#该包在新版本中，如下实现
from
sklearn
.
model_selection
import
train_test_split
X_train
,
X_test
,
y_train
,
y_test
=
train_test_split
(
X
,
y
,
random_state
=
1
)
#将数据集拆分为训练集和测试集
from
sklearn
import
tree
clf
=
tree
.
DecisionTreeClassifier
(
max_depth
=
4
)
#建树
clf
=
clf
.
fit
(
X_train
,
y_train
)
test_rec
=
X_test
.
iloc
[
1
,
:
]
clf
.
predict
(
[
test_rec
]
)
#测试集测试，交互窗口输入
'''
Ans:
Out[16]: array(['NO'], dtype=object)
'''
y_test
.
iloc
[
1
]
#调出真实结果,交互窗口输入
'''
Ans:
Out[17]: 'NO'
'''
from
sklearn
.
metrics
import
accuracy_score
rate_ac
=
accuracy_score
(
y_test
,
clf
.
predict
(
X_test
)
)
#测试模型准确率
print
(
rate_ac
)
'''
0.8205128205128205 稳定在80%左右，建树层数对准确率影响较小
'''
'''决策树可视化'''
'''
with open("lc-is.dot", 'w') as f:
f = tree.export_graphviz(clf,
out_file=f,
max_depth = 3,
impurity = True,
feature_names = list(X_train),
class_names = ['AB', 'NO'],
rounded = True,
filled= True )
'''
from
sklearn
.
tree
import
DecisionTreeClassifier
import
pydotplus
#若提示没有此包，需在cmd-Anaconda Prompt键入install pydotplus
from
IPython
.
display
import
Image
from
IPython
.
display
import
display
from
sklearn
.
tree
import
export_graphviz
#需手动下载并配置绝对路径
import
os
os
.
environ
[
"PATH"
]
+=
os
.
pathsep
+
'C:/Program Files (x86)/Graphviz2.38/bin/'
#沙雕pydotplus，配置环境变量（路径）
dot_tree
=
tree
.
export_graphviz
(
clf
,
out_file
=
None
,
feature_names
=
[
'V1'
,
'V2'
,
'V3'
,
'V4'
,
'V5'
,
'V6'
]
,
class_names
=
[
'AB'
,
'NO'
]
,
filled
=
True
,
rounded
=
True
,
special_characters
=
True
)
graph
=
pydotplus
.
graph_from_dot_data
(
dot_tree
)
img
=
Image
(
graph
.
create_png
(
)
)
graph
.
write_png
(
"out.png"
)
